This paper describes the speaker recognition system of the
NEC-TT team for the fixed training condition of the NIST
2016 speaker recognition evaluation (SRE16). Our system
is based on standard i-vector feature and Probabilistic LDA
back-end. The feature extractor employs multi-condition
training and LSTM-based voice activity detection to be ro-
bust against acoustic variability in the Call My Net Speech
Collection. The back-end includes feature normalization and
unsupervised adaptation methods to compensate mismatch
between the fixed training set (the past SREs) and the eval-
uation set. It also utilizes DNN-based gender and language
estimation to control the parameters of score calibration
and score fusion for each trial pair. Accordingly, our sys-
tem achieved 0.6192 minimum Cprimary and 0.6934 actual
Cprimary for the SRE16 development set.
各種検索
サポート
ヘルプ